神经网络模型自动分割测量颈椎MRI椎间盘及深层伸肌面积的可行性研究

Feasibility Study of artificial neural network model automating segmentation and measuring disc and deep extensor muscles on axial cervical magnetic resonance images

中文关键词:  颈椎深层伸肌  MRI测量  深度学习  智能分割  神经网络模型

英文关键词:Cervical spine  Deep extensor muscles  MRI  Deep learning  Intelligent segmentation  Neural network model

中文摘要:

  【摘要】 目的：探讨应用人工智能深度学习方法建立颈椎MRI深层伸肌自动分割与测量神经网络模型的可行性。方法：选择78例健康成年志愿者，其中男性30例，女性48例，年龄20~65岁（45.4±12.6岁）。对纳入的志愿者行3.0T颈椎MRI检查，获取符合标准的的颈椎C2/3～C6/7节段椎间盘中间层面轴位MRI图像345张。其中276张（80%）用于训练和验证以掩膜区域卷积神经网络（mask region-based convolutional neural network，Mask R-CNN）为基础建立的MRI自动分割测量模型，进行深层伸肌面积（deep extensor cross-sectional area，DCSA）、肌肉功能面积（functional cross-sectional area，FCSA）、椎间盘面积（introvertebral disc cross-sectional area，IDCSA）的自动分割与测量。剩余69张（20%）图像设定为测试集，由2名医师（人工组）手工测量其图像的DCSA、FCSA、IDCSA等，并与模型测量（模型组）的结果进行比较。采用平均交并比（mean intersection over union，MIoU）及平均像素精确度（mean pixel accuracy，MPA）评价模型分割效果，采用同类相关系数（intraclass correlation coefficients，ICC）及Bland-Altman方法比较模型组与人工组测量结果的一致性。结果：本神经网络模型最终整体MPA为0.920，整体MIoU为0.912；椎间盘、左、右侧深层伸肌的MPA分别为0.946、0.917与0.911，MIoU分别为0.934、0.908、0.899。在测试集中，人工组IDCSA为3.28±1.02cm2，左侧DCSA为2.84±1.11cm2，右侧DCSA为2.86±1.09cm2，左侧FCSA为2.19±0.89cm2，右侧FCSA为2.23±0.86cm2；模型组IDCSA为3.35±0.99cm2，左、右DCSA为3.19±1.16cm2和3.16±1.12cm2，左、右FCSA为2.49±0.99cm2、2.42±0.88cm2。模型组与人工组组间ICC值为0.852~0.914，组间ICC及Bland-Altman法测量结果显示两种测量方法一致性高。人工组与模型组测量图片平均时间为256.5±53.3s vs 0.109±0.402s，具有统计学差异（P<0.001）。结论：深度学习神经网络模型对于颈椎MRI水平位图像的深层伸肌、椎间盘组织的自动识别、分割与测量结果与人工测量一致性良好。

英文摘要:

  【Abstract】 Objectives: To develop and establish a new algorithm for the automated segmentation and measurement of deep extensor muscles on axial cervical magnetic resonance images based on deep learning. Methods: 78 adult healthy volunteers(30 males and 48 females with an average age of 45.4±12.6 years) were recruited and cervical spine magnetic resonance images were acquired using a 3.0T scanner in our hospital. A total of 345 axial T2WI MR cervical spine images through the middle of the C2/3-C6/7 disc were obtained. 276 MR images (80%) were used to train and validate a deep learning model which was developed based on Mask region-based convolutional neural network (Mask R-CNN) to segment and measure the deep extensor cross-sectional area (DCSA), functional cross-sectional area (FCSA), and intervertebral disc cross-sectional area (IDCSA) at the same level automatically. 69 images (20%) were classified as test set, the DCSA, FCSA and IDCSA of test set were manually measured by 2 surgeons for comparison with the results of model measurements. The mean intersection over union(MIoU) and mean pixel accuracy (MPA) were used to evaluate the segmentation effect of the model. Bland-Altman methods and intraclass correlation coefficients (ICCs) were used to examine the agreement between artificial measurement groups and model measurement. Results: The segmentation algorithm of this deep learning model achieved an overall MPA of 0.920 and an overall MIoU of 0.912. The MPA of IDCSA, left DCSA and right DCSA were 0.946, 0.917 and 0.911, respectively, and the MIoU were 0.934, 0.908 and 0.899, respectively. In the test set, the results of algorithm measurements of IDCSA, left and right DCSA, left and right FCSA (3.28±1.02cm2, 2.84±1.11cm2, 2.86±1.09cm2, 2.19±0.89cm2, and 2.23±0.86cm2, respectively) were similar to the results of artificial group (3.35±0.99cm2, 3.19±1.16cm2, 3.16±1.12cm2, 2.49±0.99cm2, and 2.42±0.88cm2, respectively). Interclass correlation coefficients (ICCs) were excellent for artificial measurement groups and the deep learning algorithm measurement group (0.852-0.914), and Bland-Altman plots also showed high levels of agreement. The mean artificial measurement time was 256.5±53.3s, and the mean model measurement time was 0.109±0.402s, which were statistically different(P<0.001). Conclusions: This algorithm automatically segments and measures cervical deep extensor muscles on axial MRI with comparable accuracy to spine surgeons.